Continuous Drill-Wear Scoring from Images of Drilled Holes in Melamine-Faced Chipboard Using Order-Supervised Ranking with Vision Foundation Models

In industry a drill is judged only as fit or worn; the three-class scheme with an intermediate warning state, introduced in earlier studies, smooths monitoring and lets tool replacement be planned ahead, yet hides the gradual nature of wear. This feasibility study proposes a continuous drill-wear index on a 0–100 scale, estimated from images of holes drilled in melamine-faced chipboard without per-hole wear measurements. The index is learned from two weak supervision signals: the drilling order within each drill’s life, through a pairwise ranking loss, and the wear classes, as anchors tied to the thresholds $W=0.2$ mm and $W=0.35$ mm. Vision foundation models (DINOv2, ConvNeXt V2) are fine-tuned on 8526 hole images from five drills and evaluated by leave-one-drill-out validation. On held-out drills the index follows the drilling order (profile-level Spearman correlation 0.97) and recovers the classes with 66.7–69.9% accuracy, similar to cross-entropy classifiers; its errors fall mainly towards the more worn class, and worn drills are rarely judged sharp. Conformal prediction sets reach a coverage near 90% on average but not on every drill, and an exploratory remaining-life extrapolation is at best comparable to a lifetime baseline. Larger tool batches are needed before industrial use.